fix(cognition): whole-word, marker-safe episodic keyword recall (recall quality)

[rysweet]

Summary

Improves episodic recall quality on the flat keyword scan
LibraryCognitiveMemory::search_episodes_by_keywords. It matched every keyword by
raw case-insensitive substring (content.contains(kw)), so a short clean token
embedded in the interior/suffix of an unrelated episode word spuriously recalled
off-topic episodes for its natural-language callers — e.g. decision ⊂ "indecision",
test ⊂ "latest", own ⊂ "download". This dilutes keyword recall and,
downstream, reasoner reliability.

The sibling ranked path (recall_episodes_ranked) already uses a word-boundary
gate (shares_word_prefix/tokenize_words). The flat scan was deliberately left on
substring only because its bracketed-marker callers
(memory_consolidation::reflection_lessons) depend on exact-substring match. This PR
closes that gap safely: it partitions query keywords by shape so clean tokens get
the word-boundary gate while markers keep substring — resolving the exact constraint
the substring path was preserved for, and reusing main's existing helpers (no second
matching policy).

The fix (marker-safe, reuses the established word-boundary gate)

• Clean keyword (non-empty, all alphanumeric — the shape creative_ideas
  ("meeting"/"conversation"/"decision") and tokenize_objective emit): matched at a
  word boundary via the existing shares_word_prefix (a query token must be a
  prefix of a whole content word). Inflection-tolerant, so meeting still recalls
  "meetings" and deploy still recalls "deployed" (no Fix the episodic-recall-returns-zero defect in Simard's cognitive memory. SYMPTOM: Memory preparation logs "preparation: 5 procedures, 0 episodes recalled (0 raw, 0 session-filtered)" and "prepared c #2299 "0 raw" regression). Kills
  the interior/suffix false positives.
• Phrase / bracketed marker keyword (any non-alphanumeric char — [reflect-occ=…],
  [reflect-key=…|…]): keeps legacy case-insensitive substring semantics; the
  reflection_lessons dedup and recurring-failure counting are byte-for-byte unchanged.
• Perf guard: a has_clean flag skips the word-boundary tokenization entirely on the
  marker-only path (count_recurring_failures scans with limit = u32::MAX).
• Also fixes a latent degenerate case: an empty-string keyword no longer matches every
  episode.

Merge-ready evidence

1. qa-team scenario (written, validated, run)

• New scenario: tests/qa-scenarios/whole-word-episode-recall.yaml
• gadugi-test validate -f tests/qa-scenarios/whole-word-episode-recall.yaml --strict
  → ✓ Scenario "whole-word-episode-recall" is valid (1 valid, 0 invalid)
• gadugi-test run -d tests/qa-scenarios -s whole-word-episode-recall
  → ✓ Passed: 1 ✗ Failed: 0 (drives the live-backend contract, the pure
  word_boundary_gate_tests, and the full memory_consolidation suite proving the
  exact-marker reflection_lessons dedup is unregressed).

2. Docs updated (user-facing surfaces)

• docs/architecture/cognitive-memory-library-adapter.md: updated the
  search_episodes_by_keywords behavior rows and added a "Keyword matching:
  word-boundary vs substring" subsection documenting the clean-token vs marker split.
• src/cognitive_memory/mod.rs: trait doc for search_episodes_by_keywords updated to
  describe the word-boundary + marker-safe semantics.

Changed surfaces: search_episodes_by_keywords recall behavior for clean keywords
(now word-boundary). Marker/phrase callers are behavior-preserving (substring path
unchanged).

3. quality-audit (≥3 SEEK→VALIDATE→FIX cycles, ended clean)

• Cycle 1 (SEEK): raw-substring gate admits interior/suffix embeddings → FIX:
  partition keywords into clean (word-boundary) vs raw (substring).
• Cycle 2 (SEEK): could this break the marker callers? → VALIDATE: markers carry
  non-alphanumeric chars → routed to the substring path; confirmed by the full
  memory_consolidation suite (148 pass, incl. reflection_lessons dedup and the PR-C
  recall/count e2e tests).
• Cycle 3 (SEEK): perf regression on the marker-only u32::MAX scan (needless
  tokenization) → FIX: has_clean guard skips shares_word_prefix when no clean
  keyword is present.
• Final cycle (VALIDATE, clean): reuse main's tokenize_words/shares_word_prefix
  (avoid a second, divergent matching policy); re-ran targeted suites, clippy --lib --tests -- -D warnings (clean), fmt --check (clean). Zero critical/high; zero
  medium correctness/security findings.

Local validation (all green): cognitive_memory:: (129), memory_consolidation (148),
creative_ideas (74), memory_ipc:: (68), the new
cognitive_memory::tests_whole_word_episode_recall (6), docs_integrity (4).
Pre-push gate: race-subset suite 443 passed, 0 failed and
clippy --all-targets --all-features --locked -- -D warnings (clean). Pre-commit gate:
release clippy --release --no-deps -- -D warnings (clean).

4. CI

• All pre-commit/pre-push hooks green locally (fmt, race-subset 443/443, release clippy,
  full clippy).
• Branch rebased on latest main (resolved a conflict from the concurrently-merged
  recall_episodes_ranked word-boundary work); full CI monitored on this PR and will be
  confirmed 100% green before requesting review/merge.

6. Focused diff

5 files, no unrelated edits:

• src/cognitive_memory/library_adapter.rs — partition + reuse of shares_word_prefix
• src/cognitive_memory/mod.rs — trait doc + test module registration
• src/cognitive_memory/tests_whole_word_episode_recall.rs — new integration test module
• tests/qa-scenarios/whole-word-episode-recall.yaml — new qa scenario
• docs/architecture/cognitive-memory-library-adapter.md — behavior/divergence doc

Standing-goal alignment

A durable code improvement to recall quality (cognition), extending the just-merged
word-boundary gate to the last substring recall path in a marker-safe way — consistent
with the perpetual "research and improve your own cognition" goal (durable PRs, not
snapshot docs).

[github-actions]

📊 Coverage Summary

Generated by cargo llvm-cov --workspace --summary-only (nightly, excluding test files)

Module	Lines	Covered	Coverage
Total	183336	153183	83.6%

_{Coverage data from CI run. Test files matching tests?/ are excluded from line counts.}